ATAS: Zhejiang-led team builds world-state memory framework for consistent agentic video generation
CSProfKGD · x · 2026-10-16
ATAS (Any Time, Any Space), from Zhejiang University with Monash and UC Berkeley (arXiv:2610.20571), is a world-state-centric framework for long-form agentic video generation.
Problem: passing prompts or recent frames between shots gives no shared spatial frame or time-indexed record, breaking consistency across nonconsecutive requests.
Approach:
- Compiles a screenplay into shot requests with explicit location, camera, and event-time references
- Maintains a scene memory of registered static geometry plus an event memory of dynamic observations, retrieved before generation
- Video agents share a persistent world state to revisit places and replay events from new viewpoints
Paper, code, and an interactive demo are available.
More from Multimodal
- Train Your Own LTX IC-LoRA With Grok Doing the Heavy Lifting — PurzBeats · 2026-10-16
- Hugging Face launches Open TTS Leaderboard: objective metrics to rank 8K+ open TTS models — dl_weekly · 2026-10-16
- Flux 3 Klein excels at lighting and atmosphere but fumbles living subjects — Schattenmensch · 2026-10-16
- Circus clown nails tightrope stunts in impressive Seedance 2.5 video demo — azed_ai · 2026-10-16
- ComfyUI puts up $25,000 to counter Higgsfield's broken open-source promise — PurzBeats · 2026-10-16
- Underwater brand films built from one jellyfish with an open-source Minimax template — w1kke · 2026-10-16